Muliple regression - lets try again

[statishtics]

I posted a question earlier (variable transformation) and am wondering if I didn't explain myself very well. Here's a more straighforward and paired down version of my questions .... thanks:




I'm running a multiple hierarchical regression in the social sciences. Any thoughts on the following questions:

1. For interaction effects, A X B where both are interval data ... if one variable requires transformation, how should I handle centering the data? If the variables are fairly well behaved and fit normality assumptions, is it ok to run the regression without centering (and with one variable transformed)?

2. One of these interaction terms is a scale, made up of 2 subscales, for which I want to test whether it moderates the effect between two other varialbes. If one of the subscales needs transformation, but the full scale and other subscale do not ... is it ok to run the analyes this way (i.e. 3 separate times, looking at interaction effects for each subscale and the total scale for each run)

Cheers,
M.R.

[danielkeeton]

What did you ultimately do? I am debating this same thing. Did you center the data or not?

Here is my post:

http://talkstats.com/showthread.php?p=16934#post16934

[TheAnalysisFactor]

I didn't see your first post, but can answer this one.

There are two reasons for centering predictors (that I know of)
1. to lower the correlation between a predictor and the interaction term that contains it
2. to ease interpretability of parameters

So if neither of these apply, you don't have to center the predictor. And why does the predictor require transformation? Influential outliers?

Do you mean you want to transform one subscale but not the other? That sounds like a nightmare to interpret. Are you sure you need to transform? Remember that there are no distributional assumptions for predictors.

Karen